Federated sparse representation-based anomaly detection

M. Melek

doi:10.1016/j.dsp.2025.105828

Return

Federated sparse representation-based anomaly detection

2026-01-21

PRE

M. Melek *

DOI：10.1016/j.dsp.2025.105828

Original

Original request for help

Save

Abstract

En 中文

Anomaly detection plays a vital role in industrial monitoring, IoT networks, and critical infrastructure protection. Sparse representation and dictionary learning have proved effective for this task, as they provide compact models of normal operation in which anomalies manifest as high reconstruction errors. However, traditional approaches assume centralized training, which is often infeasible due to privacy concerns, communication costs, and heterogeneous data distributions across clients. To address this gap, we propose two federated sparse representation frameworks for anomaly detection: a simple FedAvg-based and an improved method. Both approaches adapt K-SVD dictionary learning to the federated setting, enabling clients to collaboratively learn a global sparse model without sharing raw data. The improved method incorporates three mechanisms that enhance robustness under non-IID conditions, including intelligent atom assignment, contribution-aware weighted aggregation, and momentum-based updates to ensure stable convergence. Extensive experiments on synthetic and benchmark datasets demonstrate that the proposed framework outperforms centralized, local, and federated averaging baselines in terms of detection accuracy, stability, and scalability.

Keywords:

Federated learning

Sparse representation

Dictionary learning

K-SVD

Anomaly detection

Distributed signal processing